drivers/net/af_xdp/rte_eth_af_xdp.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2019 Intel Corporation.
   3  */
   4 #include <unistd.h>
   5 #include <errno.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <netinet/in.h>
   9 #include <net/if.h>
  10 #include <sys/socket.h>
  11 #include <sys/ioctl.h>
  12 #include <linux/if_ether.h>
  13 #include <linux/if_xdp.h>
  14 #include <linux/if_link.h>
  15 #include "af_xdp_deps.h"
  16 #include <bpf/xsk.h>
  17
  18 #include <rte_ethdev.h>
  19 #include <rte_ethdev_driver.h>
  20 #include <rte_ethdev_vdev.h>
  21 #include <rte_kvargs.h>
  22 #include <rte_bus_vdev.h>
  23 #include <rte_string_fns.h>
  24 #include <rte_branch_prediction.h>
  25 #include <rte_common.h>
  26 #include <rte_config.h>
  27 #include <rte_dev.h>
  28 #include <rte_eal.h>
  29 #include <rte_ether.h>
  30 #include <rte_lcore.h>
  31 #include <rte_log.h>
  32 #include <rte_memory.h>
  33 #include <rte_memzone.h>
  34 #include <rte_mbuf.h>
  35 #include <rte_malloc.h>
  36 #include <rte_ring.h>
  37
  38 #ifndef SOL_XDP
  39 #define SOL_XDP 283
  40 #endif
  41
  42 #ifndef AF_XDP
  43 #define AF_XDP 44
  44 #endif
  45
  46 #ifndef PF_XDP
  47 #define PF_XDP AF_XDP
  48 #endif
  49
  50 static int af_xdp_logtype;
  51
  52 #define AF_XDP_LOG(level, fmt, args...)                 \
  53         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  54                 "%s(): " fmt, __func__, ##args)
  55
  56 #define ETH_AF_XDP_FRAME_SIZE           XSK_UMEM__DEFAULT_FRAME_SIZE
  57 #define ETH_AF_XDP_NUM_BUFFERS          4096
  58 #define ETH_AF_XDP_DATA_HEADROOM        0
  59 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  60 #define ETH_AF_XDP_DFLT_QUEUE_IDX       0
  61
  62 #define ETH_AF_XDP_RX_BATCH_SIZE        32
  63 #define ETH_AF_XDP_TX_BATCH_SIZE        32
  64
  65 #define ETH_AF_XDP_MAX_QUEUE_PAIRS     16
  66
  67 struct xsk_umem_info {
  68         struct xsk_ring_prod fq;
  69         struct xsk_ring_cons cq;
  70         struct xsk_umem *umem;
  71         struct rte_ring *buf_ring;
  72         const struct rte_memzone *mz;
  73 };
  74
  75 struct rx_stats {
  76         uint64_t rx_pkts;
  77         uint64_t rx_bytes;
  78         uint64_t rx_dropped;
  79 };
  80
  81 struct pkt_rx_queue {
  82         struct xsk_ring_cons rx;
  83         struct xsk_umem_info *umem;
  84         struct xsk_socket *xsk;
  85         struct rte_mempool *mb_pool;
  86
  87         struct rx_stats stats;
  88
  89         struct pkt_tx_queue *pair;
  90         uint16_t queue_idx;
  91 };
  92
  93 struct tx_stats {
  94         uint64_t tx_pkts;
  95         uint64_t err_pkts;
  96         uint64_t tx_bytes;
  97 };
  98
  99 struct pkt_tx_queue {
 100         struct xsk_ring_prod tx;
 101
 102         struct tx_stats stats;
 103
 104         struct pkt_rx_queue *pair;
 105         uint16_t queue_idx;
 106 };
 107
 108 struct pmd_internals {
 109         int if_index;
 110         char if_name[IFNAMSIZ];
 111         uint16_t queue_idx;
 112         struct ether_addr eth_addr;
 113         struct xsk_umem_info *umem;
 114         struct rte_mempool *mb_pool_share;
 115
 116         struct pkt_rx_queue rx_queues[ETH_AF_XDP_MAX_QUEUE_PAIRS];
 117         struct pkt_tx_queue tx_queues[ETH_AF_XDP_MAX_QUEUE_PAIRS];
 118 };
 119
 120 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 121 #define ETH_AF_XDP_QUEUE_IDX_ARG                "queue"
 122
 123 static const char * const valid_arguments[] = {
 124         ETH_AF_XDP_IFACE_ARG,
 125         ETH_AF_XDP_QUEUE_IDX_ARG,
 126         NULL
 127 };
 128
 129 static const struct rte_eth_link pmd_link = {
 130         .link_speed = ETH_SPEED_NUM_10G,
 131         .link_duplex = ETH_LINK_FULL_DUPLEX,
 132         .link_status = ETH_LINK_DOWN,
 133         .link_autoneg = ETH_LINK_AUTONEG
 134 };
 135
 136 static inline int
 137 reserve_fill_queue(struct xsk_umem_info *umem, int reserve_size)
 138 {
 139         struct xsk_ring_prod *fq = &umem->fq;
 140         uint32_t idx;
 141         int i, ret;
 142
 143         ret = xsk_ring_prod__reserve(fq, reserve_size, &idx);
 144         if (unlikely(!ret)) {
 145                 AF_XDP_LOG(ERR, "Failed to reserve enough fq descs.\n");
 146                 return ret;
 147         }
 148
 149         for (i = 0; i < reserve_size; i++) {
 150                 __u64 *fq_addr;
 151                 void *addr = NULL;
 152                 if (rte_ring_dequeue(umem->buf_ring, &addr)) {
 153                         i--;
 154                         break;
 155                 }
 156                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 157                 *fq_addr = (uint64_t)addr;
 158         }
 159
 160         xsk_ring_prod__submit(fq, i);
 161
 162         return 0;
 163 }
 164
 165 static uint16_t
 166 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 167 {
 168         struct pkt_rx_queue *rxq = queue;
 169         struct xsk_ring_cons *rx = &rxq->rx;
 170         struct xsk_umem_info *umem = rxq->umem;
 171         struct xsk_ring_prod *fq = &umem->fq;
 172         uint32_t idx_rx = 0;
 173         uint32_t free_thresh = fq->size >> 1;
 174         struct rte_mbuf *mbufs[ETH_AF_XDP_TX_BATCH_SIZE];
 175         unsigned long dropped = 0;
 176         unsigned long rx_bytes = 0;
 177         uint16_t count = 0;
 178         int rcvd, i;
 179
 180         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 181
 182         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 183         if (rcvd == 0)
 184                 return 0;
 185
 186         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 187                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE);
 188
 189         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, rcvd) != 0))
 190                 return 0;
 191
 192         for (i = 0; i < rcvd; i++) {
 193                 const struct xdp_desc *desc;
 194                 uint64_t addr;
 195                 uint32_t len;
 196                 void *pkt;
 197
 198                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 199                 addr = desc->addr;
 200                 len = desc->len;
 201                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 202
 203                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 204                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 205                 rte_pktmbuf_data_len(mbufs[i]) = len;
 206                 rx_bytes += len;
 207                 bufs[count++] = mbufs[i];
 208
 209                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 210         }
 211
 212         xsk_ring_cons__release(rx, rcvd);
 213
 214         /* statistics */
 215         rxq->stats.rx_pkts += (rcvd - dropped);
 216         rxq->stats.rx_bytes += rx_bytes;
 217
 218         return count;
 219 }
 220
 221 static void
 222 pull_umem_cq(struct xsk_umem_info *umem, int size)
 223 {
 224         struct xsk_ring_cons *cq = &umem->cq;
 225         size_t i, n;
 226         uint32_t idx_cq = 0;
 227
 228         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 229
 230         for (i = 0; i < n; i++) {
 231                 uint64_t addr;
 232                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 233                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 234         }
 235
 236         xsk_ring_cons__release(cq, n);
 237 }
 238
 239 static void
 240 kick_tx(struct pkt_tx_queue *txq)
 241 {
 242         struct xsk_umem_info *umem = txq->pair->umem;
 243
 244         while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 245                       0, MSG_DONTWAIT) < 0) {
 246                 /* some thing unexpected */
 247                 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 248                         break;
 249
 250                 /* pull from completion queue to leave more space */
 251                 if (errno == EAGAIN)
 252                         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 253         }
 254         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 255 }
 256
 257 static uint16_t
 258 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 259 {
 260         struct pkt_tx_queue *txq = queue;
 261         struct xsk_umem_info *umem = txq->pair->umem;
 262         struct rte_mbuf *mbuf;
 263         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 264         unsigned long tx_bytes = 0;
 265         int i, valid = 0;
 266         uint32_t idx_tx;
 267
 268         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 269
 270         pull_umem_cq(umem, nb_pkts);
 271
 272         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 273                                         nb_pkts, NULL);
 274         if (nb_pkts == 0)
 275                 return 0;
 276
 277         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 278                 kick_tx(txq);
 279                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 280                 return 0;
 281         }
 282
 283         for (i = 0; i < nb_pkts; i++) {
 284                 struct xdp_desc *desc;
 285                 void *pkt;
 286                 uint32_t buf_len = ETH_AF_XDP_FRAME_SIZE
 287                                         - ETH_AF_XDP_DATA_HEADROOM;
 288                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 289                 mbuf = bufs[i];
 290                 if (mbuf->pkt_len <= buf_len) {
 291                         desc->addr = (uint64_t)addrs[valid];
 292                         desc->len = mbuf->pkt_len;
 293                         pkt = xsk_umem__get_data(umem->mz->addr,
 294                                                  desc->addr);
 295                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 296                                desc->len);
 297                         valid++;
 298                         tx_bytes += mbuf->pkt_len;
 299                 }
 300                 rte_pktmbuf_free(mbuf);
 301         }
 302
 303         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 304
 305         kick_tx(txq);
 306
 307         if (valid < nb_pkts)
 308                 rte_ring_enqueue_bulk(umem->buf_ring, &addrs[valid],
 309                                  nb_pkts - valid, NULL);
 310
 311         txq->stats.err_pkts += nb_pkts - valid;
 312         txq->stats.tx_pkts += valid;
 313         txq->stats.tx_bytes += tx_bytes;
 314
 315         return nb_pkts;
 316 }
 317
 318 static int
 319 eth_dev_start(struct rte_eth_dev *dev)
 320 {
 321         dev->data->dev_link.link_status = ETH_LINK_UP;
 322
 323         return 0;
 324 }
 325
 326 /* This function gets called when the current port gets stopped. */
 327 static void
 328 eth_dev_stop(struct rte_eth_dev *dev)
 329 {
 330         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 331 }
 332
 333 static int
 334 eth_dev_configure(struct rte_eth_dev *dev)
 335 {
 336         /* rx/tx must be paired */
 337         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 338                 return -EINVAL;
 339
 340         return 0;
 341 }
 342
 343 static void
 344 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 345 {
 346         struct pmd_internals *internals = dev->data->dev_private;
 347
 348         dev_info->if_index = internals->if_index;
 349         dev_info->max_mac_addrs = 1;
 350         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 351         dev_info->max_rx_queues = 1;
 352         dev_info->max_tx_queues = 1;
 353
 354         dev_info->min_mtu = ETHER_MIN_MTU;
 355         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 356
 357         dev_info->default_rxportconf.nb_queues = 1;
 358         dev_info->default_txportconf.nb_queues = 1;
 359         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 360         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 361 }
 362
 363 static int
 364 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 365 {
 366         struct pmd_internals *internals = dev->data->dev_private;
 367         struct xdp_statistics xdp_stats;
 368         struct pkt_rx_queue *rxq;
 369         socklen_t optlen;
 370         int i, ret;
 371
 372         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 373                 optlen = sizeof(struct xdp_statistics);
 374                 rxq = &internals->rx_queues[i];
 375                 stats->q_ipackets[i] = internals->rx_queues[i].stats.rx_pkts;
 376                 stats->q_ibytes[i] = internals->rx_queues[i].stats.rx_bytes;
 377
 378                 stats->q_opackets[i] = internals->tx_queues[i].stats.tx_pkts;
 379                 stats->q_obytes[i] = internals->tx_queues[i].stats.tx_bytes;
 380
 381                 stats->ipackets += stats->q_ipackets[i];
 382                 stats->ibytes += stats->q_ibytes[i];
 383                 stats->imissed += internals->rx_queues[i].stats.rx_dropped;
 384                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 385                                 XDP_STATISTICS, &xdp_stats, &optlen);
 386                 if (ret != 0) {
 387                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 388                         return -1;
 389                 }
 390                 stats->imissed += xdp_stats.rx_dropped;
 391
 392                 stats->opackets += stats->q_opackets[i];
 393                 stats->oerrors += internals->tx_queues[i].stats.err_pkts;
 394                 stats->obytes += stats->q_obytes[i];
 395         }
 396
 397         return 0;
 398 }
 399
 400 static void
 401 eth_stats_reset(struct rte_eth_dev *dev)
 402 {
 403         struct pmd_internals *internals = dev->data->dev_private;
 404         int i;
 405
 406         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 407                 memset(&internals->rx_queues[i].stats, 0,
 408                                         sizeof(struct rx_stats));
 409                 memset(&internals->tx_queues[i].stats, 0,
 410                                         sizeof(struct tx_stats));
 411         }
 412 }
 413
 414 static void
 415 remove_xdp_program(struct pmd_internals *internals)
 416 {
 417         uint32_t curr_prog_id = 0;
 418
 419         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 420                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 421                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 422                 return;
 423         }
 424         bpf_set_link_xdp_fd(internals->if_index, -1,
 425                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 426 }
 427
 428 static void
 429 eth_dev_close(struct rte_eth_dev *dev)
 430 {
 431         struct pmd_internals *internals = dev->data->dev_private;
 432         struct pkt_rx_queue *rxq;
 433         int i;
 434
 435         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 436                 rte_socket_id());
 437
 438         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 439                 rxq = &internals->rx_queues[i];
 440                 if (rxq->umem == NULL)
 441                         break;
 442                 xsk_socket__delete(rxq->xsk);
 443         }
 444
 445         (void)xsk_umem__delete(internals->umem->umem);
 446         remove_xdp_program(internals);
 447 }
 448
 449 static void
 450 eth_queue_release(void *q __rte_unused)
 451 {
 452 }
 453
 454 static int
 455 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 456                 int wait_to_complete __rte_unused)
 457 {
 458         return 0;
 459 }
 460
 461 static void
 462 xdp_umem_destroy(struct xsk_umem_info *umem)
 463 {
 464         rte_memzone_free(umem->mz);
 465         umem->mz = NULL;
 466
 467         rte_ring_free(umem->buf_ring);
 468         umem->buf_ring = NULL;
 469
 470         rte_free(umem);
 471         umem = NULL;
 472 }
 473
 474 static struct
 475 xsk_umem_info *xdp_umem_configure(void)
 476 {
 477         struct xsk_umem_info *umem;
 478         const struct rte_memzone *mz;
 479         struct xsk_umem_config usr_config = {
 480                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 481                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 482                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 483                 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
 484         int ret;
 485         uint64_t i;
 486
 487         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 488         if (umem == NULL) {
 489                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 490                 return NULL;
 491         }
 492
 493         umem->buf_ring = rte_ring_create("af_xdp_ring",
 494                                          ETH_AF_XDP_NUM_BUFFERS,
 495                                          rte_socket_id(),
 496                                          0x0);
 497         if (umem->buf_ring == NULL) {
 498                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 499                 goto err;
 500         }
 501
 502         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 503                 rte_ring_enqueue(umem->buf_ring,
 504                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE +
 505                                           ETH_AF_XDP_DATA_HEADROOM));
 506
 507         mz = rte_memzone_reserve_aligned("af_xdp uemem",
 508                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 509                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 510                         getpagesize());
 511         if (mz == NULL) {
 512                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 513                 goto err;
 514         }
 515
 516         ret = xsk_umem__create(&umem->umem, mz->addr,
 517                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 518                                &umem->fq, &umem->cq,
 519                                &usr_config);
 520
 521         if (ret) {
 522                 AF_XDP_LOG(ERR, "Failed to create umem");
 523                 goto err;
 524         }
 525         umem->mz = mz;
 526
 527         return umem;
 528
 529 err:
 530         xdp_umem_destroy(umem);
 531         return NULL;
 532 }
 533
 534 static int
 535 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 536               int ring_size)
 537 {
 538         struct xsk_socket_config cfg;
 539         struct pkt_tx_queue *txq = rxq->pair;
 540         int ret = 0;
 541         int reserve_size;
 542
 543         rxq->umem = xdp_umem_configure();
 544         if (rxq->umem == NULL)
 545                 return -ENOMEM;
 546
 547         cfg.rx_size = ring_size;
 548         cfg.tx_size = ring_size;
 549         cfg.libbpf_flags = 0;
 550         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 551         cfg.bind_flags = 0;
 552         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 553                         internals->queue_idx, rxq->umem->umem, &rxq->rx,
 554                         &txq->tx, &cfg);
 555         if (ret) {
 556                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 557                 goto err;
 558         }
 559
 560         reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 561         ret = reserve_fill_queue(rxq->umem, reserve_size);
 562         if (ret) {
 563                 xsk_socket__delete(rxq->xsk);
 564                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 565                 goto err;
 566         }
 567
 568         return 0;
 569
 570 err:
 571         xdp_umem_destroy(rxq->umem);
 572
 573         return ret;
 574 }
 575
 576 static void
 577 queue_reset(struct pmd_internals *internals, uint16_t queue_idx)
 578 {
 579         struct pkt_rx_queue *rxq = &internals->rx_queues[queue_idx];
 580         struct pkt_tx_queue *txq = rxq->pair;
 581
 582         memset(rxq, 0, sizeof(*rxq));
 583         memset(txq, 0, sizeof(*txq));
 584         rxq->pair = txq;
 585         txq->pair = rxq;
 586         rxq->queue_idx = queue_idx;
 587         txq->queue_idx = queue_idx;
 588 }
 589
 590 static int
 591 eth_rx_queue_setup(struct rte_eth_dev *dev,
 592                    uint16_t rx_queue_id,
 593                    uint16_t nb_rx_desc,
 594                    unsigned int socket_id __rte_unused,
 595                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 596                    struct rte_mempool *mb_pool)
 597 {
 598         struct pmd_internals *internals = dev->data->dev_private;
 599         uint32_t buf_size, data_size;
 600         struct pkt_rx_queue *rxq;
 601         int ret;
 602
 603         rxq = &internals->rx_queues[rx_queue_id];
 604         queue_reset(internals, rx_queue_id);
 605
 606         /* Now get the space available for data in the mbuf */
 607         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 608                 RTE_PKTMBUF_HEADROOM;
 609         data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 610
 611         if (data_size > buf_size) {
 612                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 613                         dev->device->name, data_size, buf_size);
 614                 ret = -ENOMEM;
 615                 goto err;
 616         }
 617
 618         rxq->mb_pool = mb_pool;
 619
 620         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 621                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 622                 ret = -EINVAL;
 623                 goto err;
 624         }
 625
 626         internals->umem = rxq->umem;
 627
 628         dev->data->rx_queues[rx_queue_id] = rxq;
 629         return 0;
 630
 631 err:
 632         queue_reset(internals, rx_queue_id);
 633         return ret;
 634 }
 635
 636 static int
 637 eth_tx_queue_setup(struct rte_eth_dev *dev,
 638                    uint16_t tx_queue_id,
 639                    uint16_t nb_tx_desc __rte_unused,
 640                    unsigned int socket_id __rte_unused,
 641                    const struct rte_eth_txconf *tx_conf __rte_unused)
 642 {
 643         struct pmd_internals *internals = dev->data->dev_private;
 644         struct pkt_tx_queue *txq;
 645
 646         txq = &internals->tx_queues[tx_queue_id];
 647
 648         dev->data->tx_queues[tx_queue_id] = txq;
 649         return 0;
 650 }
 651
 652 static int
 653 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 654 {
 655         struct pmd_internals *internals = dev->data->dev_private;
 656         struct ifreq ifr = { .ifr_mtu = mtu };
 657         int ret;
 658         int s;
 659
 660         s = socket(PF_INET, SOCK_DGRAM, 0);
 661         if (s < 0)
 662                 return -EINVAL;
 663
 664         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 665         ret = ioctl(s, SIOCSIFMTU, &ifr);
 666         close(s);
 667
 668         return (ret < 0) ? -errno : 0;
 669 }
 670
 671 static void
 672 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
 673 {
 674         struct ifreq ifr;
 675         int s;
 676
 677         s = socket(PF_INET, SOCK_DGRAM, 0);
 678         if (s < 0)
 679                 return;
 680
 681         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 682         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0)
 683                 goto out;
 684         ifr.ifr_flags &= mask;
 685         ifr.ifr_flags |= flags;
 686         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0)
 687                 goto out;
 688 out:
 689         close(s);
 690 }
 691
 692 static void
 693 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
 694 {
 695         struct pmd_internals *internals = dev->data->dev_private;
 696
 697         eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
 698 }
 699
 700 static void
 701 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
 702 {
 703         struct pmd_internals *internals = dev->data->dev_private;
 704
 705         eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
 706 }
 707
 708 static const struct eth_dev_ops ops = {
 709         .dev_start = eth_dev_start,
 710         .dev_stop = eth_dev_stop,
 711         .dev_close = eth_dev_close,
 712         .dev_configure = eth_dev_configure,
 713         .dev_infos_get = eth_dev_info,
 714         .mtu_set = eth_dev_mtu_set,
 715         .promiscuous_enable = eth_dev_promiscuous_enable,
 716         .promiscuous_disable = eth_dev_promiscuous_disable,
 717         .rx_queue_setup = eth_rx_queue_setup,
 718         .tx_queue_setup = eth_tx_queue_setup,
 719         .rx_queue_release = eth_queue_release,
 720         .tx_queue_release = eth_queue_release,
 721         .link_update = eth_link_update,
 722         .stats_get = eth_stats_get,
 723         .stats_reset = eth_stats_reset,
 724 };
 725
 726 /** parse integer from integer argument */
 727 static int
 728 parse_integer_arg(const char *key __rte_unused,
 729                   const char *value, void *extra_args)
 730 {
 731         int *i = (int *)extra_args;
 732         char *end;
 733
 734         *i = strtol(value, &end, 10);
 735         if (*i < 0) {
 736                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
 737                 return -EINVAL;
 738         }
 739
 740         return 0;
 741 }
 742
 743 /** parse name argument */
 744 static int
 745 parse_name_arg(const char *key __rte_unused,
 746                const char *value, void *extra_args)
 747 {
 748         char *name = extra_args;
 749
 750         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
 751                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
 752                            value, IFNAMSIZ);
 753                 return -EINVAL;
 754         }
 755
 756         strlcpy(name, value, IFNAMSIZ);
 757
 758         return 0;
 759 }
 760
 761 static int
 762 parse_parameters(struct rte_kvargs *kvlist,
 763                  char *if_name,
 764                  int *queue_idx)
 765 {
 766         int ret;
 767
 768         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
 769                                  &parse_name_arg, if_name);
 770         if (ret < 0)
 771                 goto free_kvlist;
 772
 773         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_IDX_ARG,
 774                                  &parse_integer_arg, queue_idx);
 775         if (ret < 0)
 776                 goto free_kvlist;
 777
 778 free_kvlist:
 779         rte_kvargs_free(kvlist);
 780         return ret;
 781 }
 782
 783 static int
 784 get_iface_info(const char *if_name,
 785                struct ether_addr *eth_addr,
 786                int *if_index)
 787 {
 788         struct ifreq ifr;
 789         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
 790
 791         if (sock < 0)
 792                 return -1;
 793
 794         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 795         if (ioctl(sock, SIOCGIFINDEX, &ifr))
 796                 goto error;
 797
 798         *if_index = ifr.ifr_ifindex;
 799
 800         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
 801                 goto error;
 802
 803         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
 804
 805         close(sock);
 806         return 0;
 807
 808 error:
 809         close(sock);
 810         return -1;
 811 }
 812
 813 static struct rte_eth_dev *
 814 init_internals(struct rte_vdev_device *dev,
 815                const char *if_name,
 816                int queue_idx)
 817 {
 818         const char *name = rte_vdev_device_name(dev);
 819         const unsigned int numa_node = dev->device.numa_node;
 820         struct pmd_internals *internals;
 821         struct rte_eth_dev *eth_dev;
 822         int ret;
 823         int i;
 824
 825         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
 826         if (internals == NULL)
 827                 return NULL;
 828
 829         internals->queue_idx = queue_idx;
 830         strlcpy(internals->if_name, if_name, IFNAMSIZ);
 831
 832         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 833                 internals->tx_queues[i].pair = &internals->rx_queues[i];
 834                 internals->rx_queues[i].pair = &internals->tx_queues[i];
 835         }
 836
 837         ret = get_iface_info(if_name, &internals->eth_addr,
 838                              &internals->if_index);
 839         if (ret)
 840                 goto err;
 841
 842         eth_dev = rte_eth_vdev_allocate(dev, 0);
 843         if (eth_dev == NULL)
 844                 goto err;
 845
 846         eth_dev->data->dev_private = internals;
 847         eth_dev->data->dev_link = pmd_link;
 848         eth_dev->data->mac_addrs = &internals->eth_addr;
 849         eth_dev->dev_ops = &ops;
 850         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
 851         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
 852
 853         return eth_dev;
 854
 855 err:
 856         rte_free(internals);
 857         return NULL;
 858 }
 859
 860 static int
 861 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
 862 {
 863         struct rte_kvargs *kvlist;
 864         char if_name[IFNAMSIZ] = {'\0'};
 865         int xsk_queue_idx = ETH_AF_XDP_DFLT_QUEUE_IDX;
 866         struct rte_eth_dev *eth_dev = NULL;
 867         const char *name;
 868
 869         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
 870                 rte_vdev_device_name(dev));
 871
 872         name = rte_vdev_device_name(dev);
 873         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
 874                 strlen(rte_vdev_device_args(dev)) == 0) {
 875                 eth_dev = rte_eth_dev_attach_secondary(name);
 876                 if (eth_dev == NULL) {
 877                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
 878                         return -EINVAL;
 879                 }
 880                 eth_dev->dev_ops = &ops;
 881                 rte_eth_dev_probing_finish(eth_dev);
 882                 return 0;
 883         }
 884
 885         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
 886         if (kvlist == NULL) {
 887                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
 888                 return -EINVAL;
 889         }
 890
 891         if (dev->device.numa_node == SOCKET_ID_ANY)
 892                 dev->device.numa_node = rte_socket_id();
 893
 894         if (parse_parameters(kvlist, if_name, &xsk_queue_idx) < 0) {
 895                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
 896                 return -EINVAL;
 897         }
 898
 899         if (strlen(if_name) == 0) {
 900                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
 901                 return -EINVAL;
 902         }
 903
 904         eth_dev = init_internals(dev, if_name, xsk_queue_idx);
 905         if (eth_dev == NULL) {
 906                 AF_XDP_LOG(ERR, "Failed to init internals\n");
 907                 return -1;
 908         }
 909
 910         rte_eth_dev_probing_finish(eth_dev);
 911
 912         return 0;
 913 }
 914
 915 static int
 916 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
 917 {
 918         struct rte_eth_dev *eth_dev = NULL;
 919         struct pmd_internals *internals;
 920
 921         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
 922                 rte_socket_id());
 923
 924         if (dev == NULL)
 925                 return -1;
 926
 927         /* find the ethdev entry */
 928         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
 929         if (eth_dev == NULL)
 930                 return -1;
 931
 932         internals = eth_dev->data->dev_private;
 933
 934         rte_ring_free(internals->umem->buf_ring);
 935         rte_memzone_free(internals->umem->mz);
 936         rte_free(internals->umem);
 937
 938         rte_eth_dev_release_port(eth_dev);
 939
 940
 941         return 0;
 942 }
 943
 944 static struct rte_vdev_driver pmd_af_xdp_drv = {
 945         .probe = rte_pmd_af_xdp_probe,
 946         .remove = rte_pmd_af_xdp_remove,
 947 };
 948
 949 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
 950 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
 951                               "iface=<string> "
 952                               "queue=<int> ");
 953
 954 RTE_INIT(af_xdp_init_log)
 955 {
 956         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
 957         if (af_xdp_logtype >= 0)
 958                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
 959 }